Research On SAR Image Target Recognition Method Based On Incremental Learning

With the gradual development of Synthetic Apertures Radar,SAR images have been widely used in military target monitoring,land resources,agricultural forestry,and earth resource exploration,etc.,which is an important part of multiple fields.And with the rapid development of computer hardware,the computer's computational power has got a quality leap,and the depth learning image recognition algorithm research has also received more and more attention.At present,the deep learning SAR image recognition algorithm is increasingly mature,but the traditional identification algorithm is a network that uses all of the data sets at one time in the training stage,and there is no way to respond to constant addition.Training data for category samples,it is difficult to make the network with the ability to continuously continue learning.However,in the real world,the target sample is often difficult to obtain one-time,so it is necessary to propose an incremental learning algorithm for SAR images.Incremental learning means that a learning system can constantly learn new knowledge from new samples,and save most of the knowledge that has been learned.With SAR image incremental learning,this thesis uses a playback incremental learning algorithm,and has launched this algorithm.Based on playback of incremental learning algorithms is to save some old category samples.When facing incremental learning tasks,use old samples to play back to alleviate the negative impact of catastrophic forgotten to the network.And this thesis proposes the use of the loss function of the reduced sample feature vendial dimension as a regular item,constructs a loss function training incremental learning network,and passes the feasibility of experimental verification algorithm and has a considerable identification accuracy.With the training data problem in the playback of the incremental learning algorithm,this thesis has studied the resolution algorithm.The incremental learning algorithm based on playback is in the face of incremental tasks,because the old sample data is only retained before,it will face the problem of data imbalance.In response to this situation,this thesis proposes a resilient incremental learning program.Its basic principle is to use the old-type sample feature vectors to weigh the distance from the center point of the sample,and the same sample sampling method is taken between different categories.Finally,this thesis verifies the feasibility of the algorithm and performs well on the identification rate of the output.